Dimensionality Reduction for Hyperspectral Data Based on Class-Aware Tensor Neighborhood Graph and Patch Alignment.

Area of Science:

• Remote Sensing
• Data Science
• Computer Vision

Background:

• Hyperspectral data analysis faces challenges like data redundancy and the curse of dimensionality.
• Effective dimensionality reduction (DR) is crucial for extracting meaningful information.
• Existing DR methods often struggle to fully utilize the spectral-spatial characteristics of hyperspectral data.

Purpose of the Study:

• To propose a novel dimensionality reduction (DR) algorithm for hyperspectral data.
• To leverage the tensor structure of hyperspectral data for enhanced analysis.
• To improve the efficiency and performance of DR algorithms by effectively utilizing spectral-spatial information.

Main Methods:

• Hyperspectral data is represented in tensor form using a window field to preserve spatial information.
• A class-aware tensor neighborhood graph is constructed using a tensor distance criterion.
• A patch alignment framework, extended to tensor space, is employed for global spectral-spatial information extraction.
• An iterative method is used to calculate the tensor subspace and obtain low-dimensional projection matrices.

Main Results:

• The proposed method simultaneously explores spectral and spatial information in hyperspectral data.
• Experimental results on three real hyperspectral datasets demonstrate superior performance compared to existing DR algorithms.
• The method requires fewer tensor training samples to achieve better performance.

Conclusions:

• The developed DR algorithm effectively addresses the curse of dimensionality in hyperspectral data.
• The tensor-based approach offers a significant improvement over traditional vector-based and tensor-based DR methods.
• This method provides a promising direction for advanced hyperspectral data analysis and feature extraction.